List of Key Words-Negative Prompt

[Luis M Ruiz]

Ok my friends. I feel like I had it with the three fingers and five legs. I decided to dig deeper on this issue and made a collection of key words that are great for removing some unwanted visual effects.  These work great depending on the image, so give these a try as you need them. I have the feeling this is list that will keep growing.

Worst Quality, Low Quality, Poorly Made Bad 3D, Lousy Bad Realistic, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, signature, cut off, draft, deformed bows, missing ear, error, Polydactyly, missing digits, extra digits, simple background, blank background, abstract background, Easy Negative, bad-hands-5, bad and mutated hands, bad hands, missing fingers, multiple limbs, interlocked fingers, ugly fingers, extra digit and hands and fingers and legs and arms, crown braid, deformed fingers, long fingers, comic, muscles, sinewy, bad anatomy, signature, monochrome, text, watermark, duplicate, bad-artist-anime, mature

[Jeff Prince]

This is a great strategy for removing qualities you don't want...

...or making hilarious cartoons of our profession.

 

 

 

[Luis M Ruiz]

Ha, ha. I like the last line, clothes from GAP.  I am going to use it.

[James Russell]

I personally think that using generic bulk negative prompts is problematic when working with text to image machine learning models.

 

 

I've seen a lot of people these days who just leave their negative prompt as "bad anatomy, deformed hands, deformed face, ugly hair", eventually filling their prompt cap with such phrases.

The biggest issue with this is how the machine learning treats the assessment of the negative phrases. 

For example "deformed hands" although it will overall reduce the finger deformity in hands, it will also generate images which have far less emphasis on "hands" overall, leading to people with their arms behind backs or out of frame.

A better case would be to, when refining the image, use positive phrases like "detailed hands" or "detailed hair" which steers the model towards a positive and detailed emphasis on these phrases.

 

If you've looked at a few of my previous posts on the topic of the current implementation of Stable Diffusion into VW2024 the above methodology, and prompts in general, can only go hand-in-hand with better control over the Stable Diffusion Generation - including primarily the Sampling Steps (crucial for a prompt like "detailed hair" as sampling steps <20 won't even get to hair detailing) and InPainting (the process of masking just an area of an image for generative change - for example you could mask just the hand of a person and then run 20 Batched Images over just those hands at a low "Creativity" / CFG and choose the best fit).

 

Stable Diffusion (Same engine as VW, Automatic1111 Interface)

Below is an example of a generated human on the right (text to image creation), and then shown on the left is the InPainting region (black) - noting this is Stable Diffusion XL in the Automatic1111 container.

The goal of this example is to show the ease in which InPainting can correct fingers (or other elements).

 

(Left: InPainting Mask, Right: Chosen Iteration of Human Figure)

 

The video below shows 50 example hand replacements. Each generation took around 7 seconds on a local machine to make.

 

50 Images created in 1 Batch for hand replacement, 40  Sample Steps with 7 CFG, 0.4 Denoising Strength (Creativity).

 

 

Overall, in my experience, prompts (both positive and negative) should be applied on a per image / intention basis, as starting with a bulk prompt eliminates huge portions of the dataset overall.

 

Or as @Luis M Ruiz rightly stated;

On 4/30/2024 at 2:50 AM, Luis M Ruiz said:

so give these a try as you need them

 

Edited May 1 by James Russell
Spelling.

[Luis M Ruiz]

I ran into this list created automatically in another AI app. I figured I'd add them to the list I am compiling but can also save these here.

Remember, experimentation is key! Try different negative prompts to see what works best for your desired outcome.

Ugly face, incorrect body proportions. bad drawing, bad anatomy, bad body shape, blurred details, awkward poses, incorrect shadows, unrealistic expressions, lack of texture, poor composition, text, logo, out of aspect ratio, body not fully visible, ugly, defects, noise, fuzzy, oversaturated, soft, blurry, out of focus, frame.

[Luis M Ruiz]

I found this link on negative prompts interesting:
https://blogs.novita.ai/untitled-6/

 

[zoomer]

Great Link.

 

But ....

 

Quote

"Now, let’s understand why negative prompts play a crucial role in stable diffusion. These prompts are essential in the image generation process, ensuring stable diffusion, realistic proportions, and high-quality images."

Quote

 

1st negative prompt

“ extra arms, extra legs, extra fingers, extra limbs,bad anatomy, bad proportions, blurry, cloned face, cropped, deformed, dehydrated, disfigured, duplicate, error, gross proportions, jpeg artifacts, long neck, low quality, low-res, malformed limbs, missing arms, missing legs, morbid, mutated hands, mutation,text, signature, ugly, username, watermark, poorly drawn hands, worst quality. ”

2nd negative prompt:

“worst quality, normal quality, low quality, low res, blurry, text, watermark, logo, banner, extra digits, cropped, jpeg artifacts, signature, username, error, sketch ,duplicate, ugly, monochrome, horror, geometry, mutation, disgusting”

 

 

I really would have expected that the AI does this itself automatically,

if not specifically wished for in positive prompts !?

 

Or does the target group prefer to get all those negative things mentioned in the negative prompt examples

in their results ?

Edited Wednesday at 04:09 PM by zoomer

[Luis M Ruiz]

I have encountered some AI apps that do include the negative prompts and many others show their field blank like the current AI Visualizer. Suggestion taken. 

[James Russell]

@zoomer negative prompts have an equal weighting as positive prompts; and as such we need to be super careful with blanketed negative prompting.

 

For example the terms 'gross proportions' and 'ugly' are not terms which should be baked into the negative steering of a Model or Checkpoint.

Imagine you're doing a Tim Burton, Guillermo del Toro or Wes Anderson inspired piece, the characters and theme you're aiming for are inherently 'grossly proportioned' and, to some, 'ugly'.

 

 

I agree with @Luis M Ruiz that suggesting these prompts, for the most part, is helpful for those learning AI Visualisation and as GUI feature overall - however I would not like to see these baked into the Model or Checkpoint in any way as I think it would immediately set a Model Bias which would limit creativity.

Edited Wednesday at 10:27 PM by James Russell

[zoomer]

“ extra arms, extra legs, extra fingers, extra limbs,bad anatomy, bad proportions, blurry, deformed, dehydrated, disfigured, duplicate, error, gross proportions, jpeg artifacts, long neck, low quality, low-res, malformed limbs, missing arms, missing legs, morbid, mutated hands, mutation,text, signature, ugly, username, watermark, poorly drawn hands, worst quality, worst quality, normal quality, low quality, low res, blurry, text, watermark, logo, banner, extra digits, cropped,

jpeg artifacts, signature, username, error, sketch ,duplicate, ugly, monochrome, horror, geometry, mutation, disgusting”

 

I just still think such results as most of these negative prompts should not be likely as a result by default,

but something you may add as positive prompts if you like to have it for some reason.

 

I mean the target group loving "jpeg artifacts" or such things in their results may be pretty small.

 

 

So an AI like in VW is more kind of "unbiased" (?)

I think I would have expected something more biased to standard human expectations.

 

For me that sounds like ask ChatGPT for a report and need to add negative prompts

for most basic things like don't use bad grammar or avoid typos .....